/vibe

/vibe command

Set a mood for all TTS speech this session. You interpret the mood and translate it into ElevenLabs expressive tags.

Usage

• /vibe banging my head against the wall — set manual vibe
• /vibe auto — return to automatic vibe detection (default)
• /vibe off — disable vibe entirely
• /vibe — show current vibe and mode

Modes

auto (default): the agent keeps the vibe current from the conversation. Every few user prompts a non-blocking reminder nudges the agent to glance at how the session is going and, if the mood has clearly shifted, set the vibe — using the same tag translation below. No deterministic classification: the agent has the whole-session context (the real success/failure signal) that a per-command hook never did.

manual: User-specified mood overrides auto-detection. The manual mood takes priority when choosing tags at stop time.

off: No vibe tags applied. Voice is neutral.

Your role: voice director

When the user sets a vibe, you translate it into 1-3 ElevenLabs expressive tags that capture the mood. This is a creative interpretation — use your judgment.

ElevenLabs eleven_v3 expressive tags are bracketed text that color delivery for ~4-5 words. The model understands:

• Emotions: [frustrated], [excited], [melancholy], [smug], [weary]
• Actions: [sighs], [laughs], [whispers], [yawns]
• Directions: [dramatic tone], [slow], [rushed]
• Scenes: [announcing a winner], [telling a secret]

Examples of your translation:

Mood / Signals	Tags you'd write
banging my head against the wall	[frustrated] [sighs]
just shipped a release	[excited]
3am and still debugging	[tired] [slow]
presenting to the board	[confident] [dramatic tone]

Keep it to 1-3 tags. Fewer is better — let the mood breathe.

When the auto reminder fires, read the mood the same way and pick tags: [happy] when things are flowing, [focused] mid-problem, [frustrated] [sighs] when stuck, [relieved] just after a fix, [weary] on a long grind. Skip the update when the mood hasn't changed.

Effect on music

If background music is playing, changing the vibe re-pools the music to the new mood — but the vibe tool does not drive playback. Instead its reply carries a music_hint: an instruction telling you to author the new pool and call the music tool yourself. The mood colors the genre; the genre never changes.

When music is off, the vibe reply has no music_hint and you do nothing music-wise — a vibe change then only sets the speaking mood.

Implementation

Use the vibe MCP tool for all writes. Use the status MCP tool for status queries.

• /vibe <mood>: Interpret the mood, choose tags, then call: vibe(mood="<mood text>", tags="<your tags>", mode="manual")
• /vibe auto: vibe(tags="", mode="auto")
• /vibe off: vibe(tags="", mode="off")
• /vibe (no argument): Call the status MCP tool and report current vibe_mode, vibe, and vibe_tags

After every vibe() write, inspect the reply for a music_hint. When it is present the reply also carries music.style (the genre now playing). Do the following — on the manual /vibe path and whenever the auto reminder makes you set the vibe:

1. Read music.style from the reply.
2. Author exactly 12 genre-accurate variations plus a base_prompt for (new mood × that style) — the mood shades the genre, it never replaces it (relaxing flamenco = slow soleá/guajira; intense flamenco = fast bulerías). Follow the authoring rules in /music.
3. Call music(mode="on", style="<that style>", base_prompt="…", variations=[… 12 …]). The daemon rotates the pool for free if it already exists, or generates it if it is new.

No music_hint (music off) → do nothing music-wise. No text output after changes — the panel confirms with the vibe shift.

Proving the chain (observability)

The subsystem appends stable [vibe-trace] events to a durable, greppable log file at ~/.punt-labs/vox/logs/vibe-trace.log. Both emitters — the mic server (vibe set / music) and the UserPromptSubmit hook (nudge fired) — are separate processes that append to this one file with atomic O_APPEND writes, so lines never interleave. Grep it to prove each link fired, or to catch a silent gap:

grep '\[vibe-trace\]' ~/.punt-labs/vox/logs/vibe-trace.log

Three event shapes, and what a matching pair proves:

• [vibe-trace] nudge fired counter=5->0 mode=auto — the auto-vibe reminder fired (from the UserPromptSubmit hook).

• [vibe-trace] vibe set mood=relaxing mode=auto music_playing=true hint_emitted=true style=flamenco — the vibe was set. hint_emitted=true means a music_hint was returned. music_playing is a tri-state: true/false from the daemon's playback status, or unknown when the status read failed (daemon down/protocol error) — an outage never reads as a false false.

• [vibe-trace] music on style=flamenco vibe=relaxing prompts=authored — the music tool re-pooled.

• Auto-vibe works when a nudge fired is followed by a vibe set … mode=auto. A nudge fired with no following vibe set = the agent ignored the nudge.

• Vibe → music works when a vibe set … music_playing=true is followed by a music on. A playing vibe set with no following music on = the agent dropped the hint.